Discarding the amplitude of the overlap for "frame_max"

Encoding the optimal tf-tradeoff for each band and then
applying it during quantization.

In cases where the coarse energy quantizer wasn't able to
go down fast enough, we could have a negative error that didn't
get clamped.

(rather than the lowest)

Apply a one-level Haar transform (or Hadamard) on each short MDCT
to nearly double the time-domain resolution.

When a transient occurs in the overlap between two frames, we
now make sure that both these frames are coded as transients

Adding one more level of band splitting so that splitting at the
PVQ encoding level is no longer necessary.

Also a bunch of comments

Now everything is done through quant_all_bands() and quant_band().

We prevent as many as possible, but they can still occur if we
manage to reach the LN limit

Uses a triangular PDF for coding the angle.

Folding now moved to alg_quant() and alg_unquant(). No more funny short block
alignment checks because bands now have to be aligned.

The main encode call no longer takes a pointer for the optional resynthesis.
It's now done with a call to celt_encode_resynthesis().